Semantic segmentation service for ice cover

The extreme natural and climatic conditions of the Arctic zone of the Russian Federation largely determine the complexity of navigation in the Arctic waters. This increases the value of obtaining accurate and timely forecasts of hydrometeorological and ice conditions. So, when making decisions when choosing a route, ship captains need to understand the state of the ice cover, namely, the ratio of different classes of ice, which determines the possibility of a vessel passing through this section. Decoding satellite images and highlighting ice classes of different structure and age in images based on the difference in their reflectivity, and as a result, the color in the image, is one of the ways to segment the ice cover. The task of decoding satellite images (semantic segmentation of the ice cover) can be solved using artificial intelligence algorithms. The service performs semantic segmentation of satellite images in accordance with 9 classes of objects: water and 8 classes of ice cover (initial types of ice, nilas, pancake, gray, gray-white, annual, drifting and solder). The service receives two satellite images (in HH and HV polarizations) of the same geographical and time reference. The result of segmentation is a single–channel image in which the areas corresponding to the same class are colored in the same color. To solve the problem of segmentation of the ice sheet, a neural network model based on the U-Net architecture was trained on 220 satellite images marked up by ice experts. In the process of developing the solution, optimal parameters of data preparation were selected (parameters of fragmentation, methods of combating class imbalance), hyperparameters of the architecture itself (number of layers, maps of hidden features) and its learning process (various loss functions, regularization methods, hyperparameters of learning were tested). As a result of experiments, the optimal option was selected, which is the basis of the service. The segmentation process is carried out in 3 stages: 1. The preprocessing of satellite images involves the conversion of sizes, resolutions and data formats. Next, the images are divided into small fragments (256*256 pixels) with an intersection equal to 224 pixels (7/8 of the fragment size). 2. The pre-trained neural network model U-Net processes all fragments obtained in the first step. The model assigns a probability distribution of belonging to the target classes to each pixel of the image. 3. The final stage of segmentation is the formation of the resulting map for all processed fragments. Since each pixel falls into several fragments during processing, we obtain a set of probability distributions, from which we select the class with the highest average probability for all predictions.